Got pure dark depth map in my own training

Question

Got pure dark depth map in my own training

Closed this issue 2 years ago · 5 comments

You know I am learning from this project but I found a strange thing. At first I generated a training dataset without 2011_09_26 in train_file.txt by accident. And therefore got pure dark depth maps in inference. But I just added 2011_09_26 in train_file.txt and it got normal. How could this happen I wonder. Thanks for your answer.

Answer 1 · 2022-08-26T01:24:56.000Z

Hi @parallelarc
I tried removing all files prefixed with 2011_09_26 and trained a model. The results in the tensorboard seem ok. I attached the revised train_files.txt for your reference.

train_files.txt

Answer 2 · 2022-08-26T06:19:28.000Z

@bolianchen Thanks，I am trying to repeat the bug

Answer 3 · 2022-08-27T03:15:05.000Z

Sorry to make you wait so long but that was really an arduous task for my computer. I used your provided train_files.txt but also got black depth maps (but not pure black). If I uses the full-loaded training file, it's ok. I am doing more experiments to see if there are any problem in my dataset or not.
Btw, how can we evaluate our models? I mean, I want to get some criterion values such like Abs Rel. I wonder if there are any of-the-shelf codes.

Answer 4 · 2022-08-28T11:48:54.000Z

Hi, @bolianchen , I decide let the bug go. The training can normally progress anyway. But I still have a question, how to train with stuct2depth format.

Answer 5 · 2022-08-30T10:17:59.000Z

Hi @parallelarc
The struct2depth format concatenates 3 adjacent frames and masks into ones. It is applied in the official tensorflow repository. We believe it induces duplicates in the training data and decide to follow how monodepth2 organizes data.